Informed Trading in Options Markets and its Information Value

Transcription

1 Informed Trading in Options Markets and its Information Value by Justin Vitanza Submitted in Partial Fulfillment of the Requirements for the Degree Doctor of Philosophy Supervised by Professor G. William Schwert and Professor Robert Ready Business Administration Simon School of Business University of Rochester Rochester, New York 2015

2 For those that believed in me when I did not ii

3 iii Biographical Sketch Justin Vitanza was born in Oxford, New York on August 7, He attended Cornell University from 2003 to 2007, and graduated with a Bachelor of Arts degree in Economics and Mathematics. He spent a brief time working for the Federal Reserve as a Research Assistant before coming to the University of Rochester in the summer of 2009, where he began his studies in Finance thanks in part to the generous funding from the Graduate School of Business Fellowship and the Robert L. And Mary L. Sproull University Fellowship. He received the Master of Science in Business Administration from the University of Rochester in His research in empirical asset pricing was conducted under the direction of Professors G. William Schwert, Robert Ready, John Long, and Robert Novy-Marx. study: The following publication was a result of work conducted during doctoral Arteta, Carlos, Steven Kamin, and Justin Vitanza, The puzzling peso, Journal of International Money and Finance, 30(8), , 2011.

4 iv Acknowledgements Thank you to Bill Schwert, from whom I have derived whatever statistical intuition I have. To Robert Ready, for an endless supply of inspiration and confidence, even when the latter was undeserved. To John Long, for keeping me honest yet allowing correct intuition to suffice. To Robert Novy-Marx, for the key insight of this paper and not settling for half-hearted explanations. Thank you also to seminar participants at the University of Rochester and various government offices throughout Washington, D.C, particularly Ron Kaniel, Michael Raith, Lenny Kostovetsky, and Paul Hanouna. To Matt Gustafson, for challenging me in this and everything else. To Ivan Ivanov, for teaching me that persistence can be as important as insight. To Mihail Velikov, for constant motivation and inspiration. Finally, and most importantly, to my wife, Bethan Lemley, for her unwaivering support and seemingly unconditional love.

5 v Abstract In this paper, I present evidence that informed traders represent a large enough portion of option market activity to impact market prices. By entering the market on the long side before positive or negative events, they drive up both open interest and ask prices, while bid prices remain relatively stable. Seeing this pattern is indicative of either positive (when found in calls) or negative (for puts) future news announcements. When conditioning on these announcements, we also see that this pattern predicts return reactions. In particular, information embedded in option prices is useful in predicting earnings surprises and reactions to mergers. My primary measure of option information content is the change in the difference between implied volatility and realized daily volatility measured over the previous month. With hindsight, this difference rises prior to positive announcements for call options, while it rises prior to negative announcements for put options. This differential behavior provides strong evidence that these assets are not redundant in practice, as is often implied by option pricing models. Further, this information constitutes a primary risk factor in equity markets, as positive announcement risk is positively related to future returns due to the procyclicality of these announcements. Efficiently utilizing this information suggests a long-short trading strategy that yields over 1.2 percent per month. This strategy also completely explains the call-put volatility spread anomaly and is robust to controls for aggregate volatility sensitivity and known metrics that purport to monitor informed trading.

6 vi Contributors and Funding Sources This work was completed under the supervision of my committee, cochaired by Robert Ready and G. William Schwert, and including John Long and Robert Novy-Marx. Funding for my time at the University of Rochester was generously provided by the Graduate School of Business Fellowship and the Robert L. And Mary L. Sproull University Fellowship.

7 vii Table of Contents 1 Introduction 1 2 Relation to Literature 8 3 Data 11 4 Theory Price Effects Volume Effects Hypotheses Pre-Earnings Announcement Volatility Changes 20 6 Trading Strategy 34 7 Price Vs. Volume 45 8 Explaining the Call-Put Volatility Spread Anomaly Time-Series Comparison Robustness Conclusion References Appendix 79

8 viii List of Tables Table 1: Summary Statistics 13 Table 2: Effect of Option Prices on Future 22 Earnings Surprises Table 3: Effect of Option Prices on Merger 28 Announcement Returns Table 4: Probits of Merger Announcements 32 Table 5: Merger Prediction Effectiveness 33 Table 6: Sorted Portfolio Returns: 39 Implied-Realized Spread Call-Put Double Sort Table 7: Sorted Portfolio Returns: 42 Implied-Realized Spread, No Earnings Effect Table 8: Sorted Portfolio Returns: 48 Double Sort with Option-to-Stock Volume Ratio Table 9: Spanning Tests: Call-Put Volatility 54 Spread vs. Implied-Realized Spread Changes (HCMHP) Table 10: Sorted Portfolio Returns: 66 Double Sort Against Liquidity Table 11: Sorted Portfolio Returns: 67 Double Sort Against Risk-Neutral Skewness Table A-1: Sorted Portfolio Returns: 79 Implied-Realized Spread/Size Double Sorts, No Earnings Effect Table A-2:Portfolio Sorts: Unconditional 80 Call-Put Double Sorts (HCMHP) Table A-3: Spanning Tests Relating HCMHP 81 to Call-Put Volatility Spreads Table A-4: Estimated Jump Risk of Long-Short Portfolios 82 Table A-4: Fama-MacBeth Regressions, No Earnings Effect 83

9 ix List of Figures Figure 1: Implied Volatility Around Earnings Announcements 25 Figure 2: Implied Volatility Around Merger Announcements 29 Figure 3: Implied Volatility Cycle Around Earnings Announcements 36 Figure 4: Time Series of Portfolio Returns 60

10 1 1 Introduction The relatively recent availability of detailed, comprehensive data for single stock options has piqued interest in using option markets to explain the crosssection of underlying asset returns. One of the more robust findings in this category is the call-put volatility spread anomaly, first documented by Bali and Hovakimian (2009). It shows that the difference between implied volatility measured on an array of call options and the same measure using put options is positively related in the cross-section to future returns. Here, I provide compelling evidence that this anomaly is caused by information spillover from options markets and an increase in positive announcement risk. That is, informed investors use option markets to profit from their information, and their activity is observable. While observable, this news is not fully accounted for in equity markets in a four-factor sense. A value-weighted long-short portfolio exploiting this factor earns over 1.2% per month, well outside of normal considerations for trading costs, especially since the sample used here is biased towards large, liquid firms. This indicates that investors require a premium to trade in stocks that have higher likelihoods of future good news, due to the procyclical nature of positive news. I show this using two distinct strains of evidence. First, I demonstrate that option open interest and implied volatilities predict future news events, such as earnings surprises and mergers. Further, the predictive power of implied volatilities is entirely concentrated in ask price data. This is consistent with informed price pressure on the long side of option markets prior to these announcements. Ex ante, it s not clear that only the ask should be affected

11 2 by informed price pressure, but it should be the case that the bid is affected no more than the ask. Consider an insider attempting to profit from their knowledge and a static order book. The first order effect is that the option supply will be taken up by informed trader, bringing the ask to the next highest supplier. If this trade does not release information to the market, then this is the only effect. However, more than likely this may trigger adjustments. In particular, sellers of options, who have the most at risk in this scenario, may increase the price of their offers, or even pull them entirely. Prospective buyers may also wish to trade on this news, but are left with little opportunity. If they react no quicker than the sellers, there is disincentive for action, as the ask price has either already or simultaneously increased and they believe that more informed traders have already acted on that market. Second, I show that these facts help us to more strongly identify the relation between volatility spreads and returns. In particular, the call-put spread anomaly is better predicted when using changes in the spread, rather than levels, and when using ask prices to calculate implied volatility, instead of the midpoint. These are precisely the components of the original anomaly that one would expect to be related to informed price pressure given the patterns observed prior to news events. I also show that the anomaly can be separately identified using either of its components (changes in call premiums and changes in put premiums). The lack of this evidence has long been a source of contention in this particular anomaly, and I believe is one of the reasons that the informed trading explanation has not been readily accepted. For example, Campbell and Petkevich

12 3 (2014) suggests that the call-put volatility anomaly is due to aggregate volatility risk sensitivity. If the spread is truly due to informed trading then we should expect that there is some component of implied volatility that will have opposing predictions for return when it is extracted from either call or put data. However, to my knowledge, no such pattern has been shown prior to this paper. The main reason for this is that prior literature has not accounted for the periodicity of volatility that arises due to scheduled quarterly announcements, a problem that does not arrive when looking at call-put spreads since the effect is more or less symmetric. An et. al. (2014) has shown that changes in call implied volatilities are positive predictors of returns, and that changes in put-implied volatilities, conditional on changes in call-implied volatilities are negative predictors. However, the former is likely picking up quarterly announcement cyclicality, and the latter is mechanically linked to the put-call volatility spread. Instead, we would ideally like to find two separate phenomena that do not depend on each other to exist. I argue that to the extent that informed trading is the cause of the predictive power of the spread, we should be able to show that such a spread is the result of two separate mechanisms: First, when call prices are higher than would be expected from volatility prices and historical premiums, we would expect future equity returns to be higher. Second, when put prices are higher than expected, we would expect lower future equity returns. The call-put spread aggregates both of these cases (using the other price to determine a baseline). If this explanation is correct, then in principle it should be possible to decompose these two effects, and discover each separately. By controlling

13 4 for other known predictors and controlling for the periodic implied volatility cycle generated by quarterly earnings announcements, I am able to isolate this component and find that this component is positively related to future returns when measured with calls, and negatively related when measured with puts. Finding this decomposition is important because it is one of the sole sources of separation between the informed investors hypothesis and other explanations. The latter typically rely on the fact that the call-put volatility spread is simply a proxy for other forms of risk. However, there is no empirical or theoretical reason to believe that the decomposition would be similarly related. For example, there is no reason to believe that firms with runups in their callimplied-realized spread have higher aggregate volatility sensitivities. However, we observe empirically that firms with a high call-put volatility spread are more sensitive. Yet, under the informed trader hypothesis, we do expect calls and puts to behave differently if we can isolate the correct component. As a result, finding this differential predictor lends credence to the informed trader explanation of the anomaly. Ideally, the base metric we would like to consider is the difference between implied volatility and average forecasted volatility. Unfortunately, the latter is not reliably available. The most basic measure is simply the lagged realized volatility, however, there are two main problems with this that need to be accounted for. First, the difference between implied and realized volatility is known to be related to the cross-section of returns. At its core, this is essentially the vol-of-vol premium (see, e.g. Bollerslev, Gibson and Zhou, 2011). To correct for this, I simply use the change in this spread as my measure

14 5 of information flow. The intuition is simple: Recognizing that the level of the implied-realized volatility spread is meaningful, I interpret changes in this level as information flow separate from this premium. In other words, I assume that this premium remains unchanged from the previous period. The advantage of this assumption is that its violation works against the main prediction of the informed trader hypothesis: If changes in this level are primarily caused by changes in the vol-of-vol premium, then we should see that this measure will have the same explanatory power in the cross-section for both call and put options. On the other hand, if changes in the level are related to information flow, then we will see that the direction of information is important, and call and put options will have opposite predictions (i.e. increases in the implied-realized volatility spread for call options will positively predict returns, while the same for put options will negatively predict returns). This paper demonstrates that after controlling for known related anomalies, we in fact see that puts and calls have differential predictions: When call prices increase more than realized volatility, future returns are high. However, when put prices increase more than realized volatility, future returns are low. Another complication is that there are predictable components to volatility. Notably, most companies issue quarterly earnings announcements, and realized (as well as implied) volatility increases during this time (Dubinsky and Johannes, 2006). While this would not necessarily bias portfolio sort results on its own (though it would greatly decrease the power), it is also a well-documented fact that earnings announcement months experience much larger returns than other months (Frazzini and Lamont (2007) estimate that

15 6 the earnings month minus non-earnings months premium is on the order of 7 percent per year). Thus, any measure of the forecasted volatility that does not somehow recognize this fact is potentially biased, and could lead to spurious conclusions. Because the amount of information expected to be released in an announcement is potentially variable, I correct for this by simply eliminating all traces of earnings announcements from the data: I only include months in which both the lagged realized and implied volatility forecast do not include an earnings announcement. The changes, then, are quarterly ones. For example, if a company has earnings announcements in March, June, etc. then the first implied-realized volatility spread is taken at January s end. The implied volatility covers the month of February, while the realized volatility is over January, neither of which contains expected earnings announcements. The second spread is taken at April s end, and the difference is the relevant metric. This does significantly reduce the power of the test: In any given month, a third of companies on average (and worse, the distribution is skewed) will be in the pool to choose from. However, despite this limitation, I still find the expected result: Changes in the implied-realized volatility spread for call options positively predict the cross section of returns while the same measure for put options is negatively related to returns. While the above description separates competing explanations for the callput volatility spread anomaly, it does not provide direct evidence that options contain information about future events. To that end, I look at two common events and show that option prices behave in ways consistent with the informed

16 7 investor hypothesis. First, I show that earnings announcement reactions are positively tied to changes in the implied-realized call volatility spread and negatively related to changes in the implied-realized put volatility spread. Second, I show that call implied volatility runs up prior to merger announcements for targets. These facts, taken together with the portfolio sort analyses, suggest that calls and puts both have predictive power in explaining the cross section of equity returns, and that this explanatory power comes from relatively informed traders making bets on future announcements.

17 8 2 Relation to Literature The analysis presented here is primarily an extension and explanation of the findings in Bali and Hovakimian (2009). In it, the authors demonstrate that the difference in implied volatilities between matching call and put options is positively related to future stock returns at the monthly level. While there are several potential explanations for this anomaly presented, they can be separated into two major categories: Those that see this measure as informed price pressure, and those that see it as a proxy for undiversifiable risk. The two most recent papers examining this basic dichotomy come down on opposite sides of this question. An et. al. (2014) shows that changes in the callimplied volatility are positively related to future returns, and that put-implied changes, conditional on call-implied changes, are negatively related. They view this as evidence that implied volatility from puts and calls contain differential information about future returns, and that it therefore reflects informed price pressure on these assets. Campbell and Petkevich (2013) take the opposite view. They show that if you control for aggregate volatility sensitivity, the alpha generated from a long-short portfolio sorted on the call-put implied volatility spread becomes statistically insignificant. My analysis suggests that both of these explanations are partially correct. First, I demonstrate directly by looking at various corporate events that the call-put volatility spread is indicative of future news, and the stock returns attributed to the release of that news. This predictive power is also signed, meaning that the spread runs up before positive news but down before negative news. Further, it can be isolated and found in either call- or put-implied volatilities unconditional

18 9 on the other, and it appears that only options in which the long side has substantial upside risk for the type of news announcement being analyzed has predictive power. This signifies that the spread has information not contained by the rest of the market on short time horizons. However, the spread also contains information about aggregate volatility sensitivity. In particular, I show that the spread has predictive power over returns at a longer duration than it has for events. Further, the predictive power is lessened (though not eliminated) by controlling for aggregate volatility sensitivity. There is also a longer string of literature examining the relation between option and stock markets more generally, and in particular how informed traders participate in these markets when both are available to them. Perhaps the foundation of the latter type is Easley, O Hara and Srinivas (1998). It develops a multimarket equilibrium model in which traders can be active in both markets, and predicts that information in such a market should flow from the options markets to the underlying markets. In particular, signed volume trading should be indicative of future price movements in equity markets. Pan and Poteshman (2006) test this relatively directly and indeed find that firms with more newly initiated positions in calls relative to those in puts is a positive predictor of future returns. Their data is proprietary, indicating that this does not represent a market inefficiency, but a case in which traders with nonpublic data can outperform those without it. Slightly later, Roll, Schwartz and Subrahmanyam (2010) showed that a publicly available version of this concept - the option-to-stock volume ratio - negatively predicts returns surrounding earnings announcements. Johnson and

19 10 So (2012) then showed that it predicts future returns more generally. Unfortunately, the option-to-stock volume is a much weaker signal than that used by Pan and Poteshman (2006). This is because the option-to-stock volume ratio relies theoretically on the fact that short sale constraints make option markets more attractive relative to equity markets precisely when bad news is more likely. Thus, the option-to-stock volume ratio is a portent of bad news, even though the measure uses both call and put options in its construction. Johnson and So (2012) also demonstrates that the predictability of this measure is positively related to various measure of short sale constraints. While this volume analysis is related to my analysis in this paper, it turns out that these measures are complementary and not interchangeable. Using unsigned but publicly available data, Blau and Nguyen (2014) show that the put-call volume ratios similar to those of Pan and Poteshman (2006) are only predictive of future returns at relatively short levels (less than weekly). However, the option-to-stock volumes of Johnson and So (2012) has predictive persistence of up to six weeks. This duration is similar to what is seen from the call-put implied volatility spread examined here and in Bali and Hovakimian (2009). As a result, one might be concerned that these two measures are observing the same phenomenon. While there does appear to be some redundancy, neither measure can explain the predictive power of the other. Thus, we expect that using volume and implied volatility information can be joint predictors of informed trading, and not merely substitutes.

20 11 3 Data Implied volatility (hereafter, ˆσ) is taken from OptionMetrics, spanning from January 1, 1996 to December 31, The implied volatility typically used is that of the call or put option closest to at-the-money with the closest expiration, and is the volatility implied by the Black-Scholes options pricing model at end of day. ˆσ is adjusted by OptionMetrics to account for expected dividend payments and early exercise premia. Unfortunately the method used for these adjustments is proprietary. In monthly regressions, the value on the last trading day of the month is taken. Changes in implied volatility, ˆσ, are the difference between implied volatility the current month and that in the previous month. Open Interest is the number of outstanding contracts for the option contract used to construct implied volatilities. Stock returns, bid and ask prices, and market equity are taken from CRSP from the WRDS database. Fama-French factors are taken from the Kenneth French data library. Call and put option volumes are taken from the publicly available dataset from the CBOE. Merger dates are from SDC and include all announcements in US dollars. Mergers are linked to CRSP Permnos through the CRSP linking tool in WRDS. The binary Acquirer variable takes the value of 1 if, during the month, a merger is announced with that firm as the acquirer. T arget takes the value of 1 if, during the month, a merger is announced with the firm as a target and not as an acquirer. Following Schwert (1996), I only include announcements in which no prior takeover announcement has been issued in the previous year. Earnings Announcement dates are taken from Compusat Quarterly (RDQ).

21 12 The following variables are taken from the Compustat Annual file: To convert the variables to monthly, the values are repeated for months after the initial data date. Size is the log book assets (AT). Debt is total long-term debt scaled by assets (DLTT/AT). Leverage is total long-term debt scaled by common equity (DLTT/CEQ). Earnings is retained earnings scaled by assets (RE/AT). All variables are winsorized at the 1% level. Coverage of the OptionMetrics database goes from just over half of the CRSP/Compustat merged database in 1996 to nearly three-quarters of it in However, the characteristic makeup of the firms over this sample doesn t appear to change very much with the coverage. In general, the subsample used in this paper is a bit larger, especially in the early period where coverage is sparser. However, as coverage increases, the size of firm appears to converge towards that of the full sample. However, this subsample also contains firms with significantly higher leverage, and somewhat higher debt, and this persists throughout the sample despite the increase in coverage.

22 13 Table 1: Summary Statistics Table 1 shows summary statistics for some key variables in both the full CRSP/Compustat Annual File and the subset that has been matched to OptionMetrics. Size is the log book assets (AT). Earnings is retained earnings scaled by assets (RE/AT). M/B is Market Value (CEQ) scaled by total assets (AT). Leverage is total long-term debt scaled by common equity (DLTT/CEQ). Debt is total long-term debt scaled by assets (DLTT/AT). Coverage is the percent of observation in the subsample compared to the full sample. All variables are winsorized at the 1% level. Panel A: Full CRSP/Compustat Merged: Annual File Year Earnings Size M/B Leverage Debt Panel B: CRSP/Compustat/OptionMetrics Merged Year Earnings Size M/B Leverage Debt Coverage

23 14 4 Theory 4.1 Price Effects Easley, O Hara and Srinivas (1998) create a simple model in which investors are able to simultaneously trade in both option and stock markets. The model is a static one in which nature determines whether there is an information event, and whether it is a good or bad signal. At this point, informed and uninformed traders both take actions. Uninformed traders, which are all traders in the state where no information occurs, have a propensity to trade on all sides. That is, they both buy and sell stocks, call options, and put options. If there is an information event, however, informed traders take one of three actions corresponding to the type of information they receive. For example, in the event of good news, informed investors will either buy stocks, buy calls, or sell puts. Again, they split among these three alternatives with certain propensities. There is also a risk-neutral market maker with a zero profit condition. While the paper itself does not spend much time exploring this fact, it does solve for the bid and ask prices of all of the available investments. The four prices related to options are especially informative. For both puts and calls, the bid price is related to the informed ratio of option sellers while the ask price is related to the informed ratio of option buyers. For example, during a good news event, there will be both informed and uninformed investors attempting to sell puts. The larger the fraction of these investors are informed, the lower the bid price that the market maker (here, a stand-in for the equilibrium

24 15 price) is willing to buy at. Similarly, the larger the fraction of call buyers are informed, the higher the ask price the market maker will demand. This makes intuitive sense: If you are attempting to buy a call, you must take into account the probability that the other side of the trade is informed. The higher that probability, the higher the chance that you are entering a losing trade, and should thus appropriate discount your willingness to pay. Similarly, if you would like to sell a call, there is some probability that the population you are selling to has positive information. The important takeaway here is that informed traders long in option markets affect ask prices while informed traders short in option markets affect bid prices. Because of this, the extent to which we should observe deviations on market prices in bid and asks due to corporate announcements is going to depend on the frequency with which informed traders invest on the long or short side of option markets, or at the very least on the market s belief about those frequencies. Unfortunately, here there is very little theory to guide us. Intuitively, being long in an option market should seem the more attractive option because there is higher leverage to exploit your information. However, there may be other considerations such as ease of detection that could counterbalance the incentive effects of leverage. 4.2 Volume Effects A consequence of the multi-market model of Easy, O Hara and Srinivas (1998) is that any positive signal trade (buying calls or selling puts) will result in increased equilibrium bid and ask prices in stocks and calls, while decreas-

25 16 ing these prices in puts. Negative signal trades will have the opposite effect. However, the extent to which it will affect these prices is dependent on the likelihood that the trade is informed. So by the same logic by which we deduced that bid and ask prices are affected by the proportion of informed traders, we may also infer that the predictability of future prices. Since these two variables are shifted by the same effect, we can also hypothesize that bid and ask prices will predict future prices, absent any other correlation. Pan and Poteshman (2006) use signed option volume as another way to measure informed trades. Their proprietary data set allows them to differentiate between trades initiated by buyers and sellers. Ideally, we would want to ensure that the price effect predicted in the previous section and shown to be found in later sections is not better explained by directly observing volume. Unfortunately, this data is not available for my current analysis. However, other papers have shown that the persistence of this phenomenon is much shorter (less than weekly) than the effect I observe on prices (monthly). However, Johnson and So (2012) show that option-to-stock ratios are not short-lived, and in fact predict future returns for up to six weeks. The use of option-to-stock volume ratios is not a direct test of a specific model, but instead a reasoned argument: Price discovery in stocks can be hampered by short sale constraints in the event of bad news. However, option markets, by their symmetric nature, allow for this price discovery to happen with fewer constraints. As a result, activity in option markets (relative to activity in the underlying) should be a portent of bad news. This is shown to in fact be the case, particularly in stocks where short sale constraints are strict.

26 Hypotheses Based on previous theoretical and empirical work, I would expect that the call-put volatility spread s predictive power for future returns can be tied to informed trading in options markets prior to public release of that information. This section summarizes the series of tests that I will perform, relating the expected results to previous established work in the field, and discerning when these results would differ from alternative explanations, where appropriate. First, and most obviously, I should expect to see that call-put volatility spreads should actually be a predictor of corporate announcements. If, as is my claim, the call-put volatility spread is a proxy for the amount of insider information embedded in option prices, then one would expect that when looking at information releases, ex post, we will see abnormal behavior in the spread that coincides with the direction of the future news. To that end, I look at three different types of announcements. First, I examine activity surrounding earnings announcements. This announcement is a predictable one in frequency but whose content is unknown. This is perhaps the best test of the informed trader hypothesis because it has differential predictions from alternatives. Under the informed trader hypothesis, positive call-put volatility spreads are indicative of positive future information. Thus, if an earnings announcement is good, we should expect an increase in this spread prior to the announcement date. Similarly, the measure should fall before bad news. However, under a hypothesis in which the coming announcement is the cause of the spread not by content but by existence, then we should expect to see the same pattern in the spread before either good or bad announcements. The

27 18 same is true if we expect that the spread is unrelated to the announcement. In fact, the only possible conflating explanation would be one that is correlated with the content of the news. I also examine merger activity, which is typically associated with large positive returns for the target in the deal. Again, I would expect to see a rise in the spread prior to these announcements. This provides an alternative, unexpected announcement to ensure that the results are not a part of a cyclical pattern surrounding predictable announcements. Second, I investigate the extent to which the predictive power of the spread is concentrated in ask or bid prices, and to a lesser extent in call or put options. The theory and intuition of the multi-market model of Easley, O Hara and Srinivas (1998) puts a clearing price for both bids and asks on both calls and puts for a risk-neutral market maker. It shows that the price is related to the likelihood of informed trading occurring in that market, such that a ask (bid) price should be higher (lower) when there is more informed trading in long (short) option markets. While I believe that informed trading is most likely to occur in long option markets due to the leverage of the asset, the extent to which informed traders participate in any of these markets is ultimately an empirical question. Pan and Poteshman (2006) suggests that this is true. In looking at signed option trades (that is, whether they are buyer- or seller-initiated), they find that call-put volume ratios are only predictive when both volumes contain only buyer-initiated trades. This suggests that the information content of volume is concentrated in the long side: Traders with more information will tend to trade long in the appropriate market. Thus, I run future tests separately using bid and ask prices to determine whether

28 19 we can differentiate the extent to which this trading is concentrated on the long or short side of option markets. To be consistent with previous work, it should be true that most of the predictive power of the spread is contained in ask prices.

29 5 Pre-Earnings Announcement Volatility Changes 20 Earnings announcements provide a regular event which frequently contain material information to the investor base that is known to insiders. This makes them excellent candidates for evaluating the likelihood of option prices reflecting inside knowledge. However, they also have a few less desirable qualities. First, because equity returns are more volatile around an announcement, the implied volatility runs up prior to earnings announcements. There is no theoretical reason, however, that this should dissimilarly affect puts and calls. Looking at the put-call volatility spread around announcement, then, we would not necessarily expect any abnormalities or jumps on or around the event date, particularly since this date is known well in advance, it is merely the content that is unknown to the public. Even with a general insider trading hypothesis, it is unclear what we should expect to see around announcement as there are multiple ways that information could be entering the market. However, I am positing a particular mechanism that has very particular consequences, which end up being borne in the data. I imagine a group of informed traders with superior information about, say, a positive upcoming earnings announcement. These investors act on this information by buying call options with market orders. Assuming that the other investors take no action in response to this event, bid prices would remain unchanged (again, I am assuming that all other orders have cleared and that there are no new orders in response to this). However, ask prices would rise: Presumably there is a supply curve for options at any given time represented by the order book. If the new order is large enough to take all the supply at the current ask, then

30 21 the ask must increase to the next level of supply. Of course, there should be secondary information effects. Once the orders come in, if they are known to be informed, new bids should come in, driving up the bid price. However, if it is not known that the new order is by informed traders, other participants may make no significant inference from the trades. Regardless of whether it is known or not, the secondary effect should be no larger than the primary, and thus I would expect that the put-call volatility spread should have more predictive power when calculated using ask prices. Table 2 shows that this prediction bears out. I measure the earning surprise by calculating cumulative abnormal returns around the announcement window. These returns are compounded over the window after subtracting the return of the CRSP value-weighted index. The measure of insider information is the change in the implied-realized volatility spread, calculated using either call or put data at either the bid or the ask price. For all measures, the difference is taken from one calendar month prior to the announcement to 4 trading days prior to the announcement in order to prevent any overlap with the CAR measure. The realized portion of volatility is calculated as the standard deviation of the daily returns for the previous calendar month. If changes in the implied-realized volatility spread are partially composed of insider trading, then we would expect that component to be correlated with the level of the surprise. For clarity, I would note that this regression is signed. That is, changes in the spread do not predict only the magnitude of earnings surprises, but also the direction.

31 22 Table 2: Effect of Option Prices on Future Earnings Surprises Table 2 shows the results of panel regressions of earnings surprises on changes in implied-realized volatility spread. The sample consists of all firms with quarterly earnings announcement dates in Compustat and option data in OptionMetrics from January 1996 to December Each regression has, as its dependent variable, the cumulative abnormal return (CAR), calculated as the cumulative difference between equity returns and the CRSP value-weighted index for period spanning from 3 trading days prior to the earnings announcement to 3 trading days after the announcement. (ˆσ σ), is equal to the level change in the difference between the Black-Scholes option-implied volatility and the standard deviation of one month of daily returns. Regressions (1) and (3) use bid prices to calculate implied volatility from call and put options, respectively. Regressions (2) and (4) use ask prices. All implied volatilities are taken from options with the shortest expiration that is closest to at-the-money. All other controls are also lagged and include size (log of market capitalization), debt (debt-to-assets ratio), leverage (debt-to-equity ratio, in percent), and earnings (earnings-to-assets ratio). Standard errors are in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% significance, respectively. Standard errors are Huber-White-sandwich heteroskedasticity- and autocorrelation-consistent standard errors. All regressions include month fixed effects. Call Put Bid Ask Bid Ask Const (.0126) (.0113) (.0130) (.0142) (ˆσ σ) (.0059) (.0063) (.0054) (.0051) Size (.102) (.112) (.130) (.125) Debt (3.80) (3.99) (4.11) (4.91) Leverage (1.10) (1.02) (0.99) (1.07) Earnings (0.875) (0.874) (0.877) (0.875) R N = 39234

32 23 This regression shows that the variable of interest is significant only if the implied volatility is measured using the ask price. Further, the sign of the prediction flips when we use put options instead of calls. This indicates that, at least before earnings announcements, there is a component of implied volatility that is differentially predictive for calls and puts. To my knowledge, this is the first direct evidence that such a component exists. While there is ever-growing evidence that options are not truly redundant assets due to the option trading process, there is little to no evidence that this process differentially affects puts and calls. To provide further evidence, I examine the event-conditional time series of call-put volatility differentials. Figure 1 plots the difference between call and put option-implied volatilities in event time, where the implied volatility is measured by nearest expiration at-the-money call/put pairs. Positive values indicate that the call implied volatility is higher than that of the put, which seems to be the case on average. Time zero indicates a specific kind of earnings announcement: In Panel A, time zero indicates an earnings announcement that experienced a CAR in the highest tercile of the sample, while in Panel B, time zero is the date of an earnings announcement with a CAR in the bottom tercile. To be in the top or bottom tercile is to be in that range for a particular year. In other words, the top tercile is not more heavily weighted by announcements in overall good years, but one third of the announcements from each year are represented in that sample. This separates the sample by announcement quality relative to their peers. Thus, Panel A is the subset of firms with good earnings announcements

33 24 while Panel B are those with bad news. Each plot contains two lines. The main line that shows the predicted behavior (increases around announcements for the good sample and decreases for the bad sample), represents the callput volatility spread measured using ask prices. The second, essentially flat, line represents the same metric using bid prices as the calculation point. As you can see, prior to a positive earnings announcement, there is a noticeable runup in the call-put volatility spread when measured using ask prices but not when measured using bid prices. This indicates that there is more positive price pressure on calls than on puts around positive earnings news. However, prior to a bad announcement, there is a very pronounced drop in the spread that actually persists for some time after the announcement. That this does not show up using bid prices is consistent with the idea of informed traders consuming option supply without affecting, on average, other market behavior, as described in the previous section.

34 25 Figure 1: Implied volatility around earnings announcements Figure 1 plots series in event time, where t = 0 represents an earnings announcement. Eash series spans from 50 trading days prior to the announcement date to 50 trading days after the announcement. Each series represents the difference between implied volatity taken from call options and that of put options. The solid line uses ask prices to calculate implied volatilities while the dotted line uses bid prices. Positive values suggest that the call implied volatility is higher than that of the put. Both the call and put volatilities are taken from nearest expiration, closest to at-the-money options. Panel A uses announcements in the highest tercile of cumulative abnormal returns (CAR) from each year. Panel B uses announcements in the lowest CAR tercile. CAR is measured as the cumulative difference between equity returns and the CRSP value-weighted index for period spanning from 3 trading days prior to the earnings announcement to 3 trading days after the announcement. Panel A: Call-put implied volatility spreads surrounding good earnings surprises. Panel B: Call-put implied volatility spreads surrounding bad earnings surprises.

35 26 The evidence provided indicates that changes in the implied-realized or call-put volatility spreads are significant predictors of earnings surprises. Since this same metric also predicts returns, this can mean one of two things. My proposed explanation is that the predictive power of option prices is the result of insiders affecting the market clearing price through increased trading. That this then translates to abnormal returns in the equity market must mean that this is unable to be arbitraged. This could be due to the noisiness of the signal or the lack of awareness by market participants. In future revisions, I plan to look at short-term trading strategies that exploit this pattern in bid and ask option-implied volatilities to see if it is possible to generate significant returns that should be arbitraged away with proper foresight by markets. Since earnings announcement dates are known ahead of time, it seems unlikely that there are events that lead to earnings surprises and also this volatility behavior in ask prices only. However, it could be the case that the informed trading surrounding earnings announcements and other corporate events are only a small portion of the future equity returns predicted by call-put volatility spreads. A similar set of empirical facts hold true for mergers. A more thorough treatment of this particular case can be find in Augustin, Brenner, and Subrahmanyam (2014). However, it is important to note that the same patterns observed before the expected event, earnings announcements, as in the unexpected event, mergers. Also important to note is that, just as in the earnings announcement sample, the predictive power of implied volatilities is concentrated purely in the ask price, and not the bid price, a fact not covered in the aforementioned analysis, which primarily uses standardized options which

36 27 prevent bid-ask differential analysis. Previously, I showed that changes in the implied realized spread have predictive power for returns surrounding earnings announcements. This generally remains true for mergers as well, though the coefficient of interest is insignificant, both economically and statistically, when it is measured using put options. While there are certainly cases in which mergers are negatively received poorly by the market, it is not clear that these will be known ahead of time (else, why would the managers make such a decision), so it is not surprising that put options do not have predictive power for returns, at least not under the specification designed to proxy for informed price pressure. Table 3 repeats the exercise done in Table 2 but with mergers being the event in question rather than earnings announcements. Again, the coefficient on the change in the implied-realized spread is only significant when calculated using ask prices, though now it is additionally only significant when measured using call options. Similarly, Figure 2 plots changes in the call-put volatility spread in event time around merger announcements for takeovers. As in the case of earnings announcements, we see a sharp rise in this measure prior to the announcement, and it vanishes nearly immediately. Again, it exists only when calculating volatilities from ask prices.

37 28 Table 3: Effect of Option Prices on Merger Announcement Returns Table 3 shows the results of panel regressions of cumulative abnormal returns (CAR) around target merger announcements on changes in the implied-realized volatility spread. The sample consists of all firms with quarterly earnings announcement dates in Compustat and option data in OptionMetrics from January 1996 to December Each regression has, as its dependent variable, the CAR, calculated as the cumulative difference between equity returns and the CRSP value-weighted index for period spanning from 3 trading days prior to the merger announcement to 3 trading days after the announcement. (ˆσ σ), is equal to the level change in the difference between the Black-Scholes option-implied volatility and the standard deviation of one month of daily returns. Regressions (1) and (3) use bid prices to calculate implied volatility from call and put options, respectively. Regressions (2) and (4) use ask prices. All implied volatilities are taken from options with the shortest expiration that is closest to at-the-money. All other controls are also lagged and include size (log of market capitalization), debt (debt-to-assets ratio), leverage (debt-to-equity ratio, in percent), and earnings (earnings-to-assets ratio). Standard errors are in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% significance, respectively. Standard errors are Huber-White-sandwich heteroskedasticity- and autocorrelation-consistent standard errors. All regressions include month fixed effects. Call Put Bid Ask Bid Ask Const (.0149) (.0142) (.0152) (.0139) (ˆσ σ) (.0106) (.0098) (.0093) (.0084) Size (.122) (.134) (.161) (.150) Debt (5.06) (4.92) (4.99) (5.31) Leverage (1.65) (1.52) (1.59) (1.67) Earnings (1.11) (1.12) (1.12) (1.11) R N = 7223

38 29 Figure 2: Implied volatility around merger announcements Figure 2 plots series in event time, where t = 0 represents a merger announcement. Eash series spans from 100 trading days prior to the announcement date to 100 trading days after the announcement. Each series represents the difference between implied volatity taken from call options and that of put options. In Panel A, the solid blue line uses ask prices to calculate implied volatilities while the dotted green line uses bid prices. Positive values suggest that the call implied volatility is higher than that of the put. Both the call and put volatilities are taken from nearest expiration, closest to at-the-money options. Panel B plots the put-call volatility spread uses ask prices in both series. The solid blue line is the same as that in Panel A. The dotted green line instead uses options with an expiration of between 30 and 90 calendar days. Panel A: Call-put implied volatility spread surrounding good earnings surprises. Panel B: Call-put implied volatility spread surrounding bad earnings surprises.

39 30 What is interesting particularly about mergers is that the runup in the implied volatility start at drastically different times for different durations of options. The exact magnitude of a quarterly earnings announcement may not be well-known months prior to release, but merger talks are likely kept secret for extended periods of time. As a result, it is possible to trade in much longer-term options than would be prudent, even for an insider, around earnings data. In all previous analyses, volatilities were taken from nearestterm options, so the time to expiration is typically less than a month. We also see that the runup in spreads also occurs well within a month, usually around 10 trading days beforehand. However, if we look at longer-term options, we see that the runup happens on a much longer and smoother basis for mergers. In Figure 2, we see that when looking at options with between one and three months to expiration, the runup occurs a full two months prior to the actual announcement date. This indicates a precision in announcement date windows that is somewhat remarkable for investors without inside information. Unlike earnings announcements, there is an additional dimension to merger announcements. Namely, they are unpredictable in nature. While there may certainly be rumors and hints that a company is looking to be taken over, there is certainly no guarantee of when or where that may take place. As a result, we should also be interested as to whether or not these measures of informed price pressure can actually predict the announcement of mergers, and not just the reaction conditional on announcement. Table 4 conducts a probit regression with the dependent variable being equal to one if the firm announces that it has received a takeover bid during

40 31 that month, and zero otherwise. I regress this on the change in the impliedrealized spread for call and put options, or alternatively in changes in the call-put volatility spread. Again, we find statistically significant predictive power in changes in both the call implied-realized spread and the call-put volatility spread, and again only when using ask prices. Table 5 briefly examines the difference in Pseudo-R 2 each variable makes to determine the predictive power of implied-realized spread changes compared to other measures used to predict mergers in previous literature. While all of these variables are poor predictors (which they should be or else mergers wouldn t be very surprising), spread changes appear to be on par with the best variables available. However, only Leverage has an increase in Pseudo-R 2 significantly different from that of changes in the implied-realized spread.

41 32 Table 4: Probits of Merger Announcements Each probit has, as its dependent variable, an indicator that takes the value of 1 if, during the month, there is merger announcement in which firm is the target, and has not been a target in the previous year. The independent variable of interest, (ˆσ σ), is equal to the level change in the difference between the Black-Scholes option-implied volatility taken from nearest-expiration at-the-money options and the standard deviation of one month of daily returns. (ˆσ 1,2 σ) 1 is the same change except that it is lagged and uses options that are between one and two months from expiration. Standard errors are in parentheses. *, **, and *** indicate significance at the 10%, 5%, and 1% significance, respectively. r 1,0 is the previous months return. Other controls are Size, the natural log of total assets, Debt, total debt scaled by total assets, Leverage, Total Debt scaled by common equity, and Earnings, retained earnings scaled by assets. All controls take the last reported value from Compustat Annual. Controls are lagged by one month. Regressions differ only in the inclusion of the two variables of interest and the way implied volatility is calculated. Columns (2) through (5) use call options to calculate volatility, with column (2) using the bid price and columns (3) through (5) using the ask. Columns (6) and (7) use put prices, with column (6) using the bid and (7) the ask. Call Put Bid Ask Bid Ask (ˆσ σ) (.083) (.069) (.094) (.099) (.089).186 (ˆσ 1,2 σ) (.083) (.102).177 r ,0 (.053) (.053) (.053) (.053) (.071) (.052) (.053) Size (.007) (.007) (.008) (.008) (.008) (.007) (.007) Debt (.050) (.050) (.051) (.050) (.050) (.050) (.050) Leverage.009*.009*.009*.009*.009*.009*.009* Earnings (.005) (.005) (.005) (.005) (.005) (.005) (.005) (.013) (.013) (.013) (.013) (.013) (.013) (.013) Pseudo-R N = 369,829

42 33 Table 5: Merger Prediction Effectiveness This table presents the difference in the Pseudo-R 2 between the full probit prediction model in column (3) of Table 4 and that in which the listed variable is removed. (ˆσ σ), is equal to the level change in the difference between the Black-Scholes option-implied volatility taken from nearest-expiration atthe-money options and the standard deviation of one month of daily returns. r 1,0 is the previous months return. Size is the natural log of total assets. Debt is total debt scaled by total assets. Leverage is total debt scaled by common equity, and Earnings is retained earnings scaled by assets. All variables take the last reported value from Compustat Annual whose announcement date is before the first of the month over which the dependent variable, whether a firm is a merger target, is taken. (ˆσ σ) r 1,0 Size Debt Leverage Earnings

43 34 6 Trading Strategy The call-put implied volatility spread has been demonstrated to be predictive in the cross-section of equity returns. However, there is little direct evidence as to why this anomaly exists. My contention is that it arises due to informed trades in relatively (compared to the underlyings) illiquid markets. This illiquidity causes demand and supply for options to be less than perfectly elastic, so that a relatively small group of traders with informed guesses will impact the clearing price. This, in turn, signals increased risk of positive (or negative) announcement risk. It is not my contention equity markets are unaware of this increased risk, but that this then becomes priced, since good news is correlated with good states of the world. To test this theory, I separate the call-put spread into two components. It has previously been documented that simply sorting on implied volatilities (either of puts or calls) has no predictive power for future returns. However, I would not expect, under an informed trader hypothesis, that the level of the volatility would be a particularly strong indicator. What should have predictive power is the clearing price relative to some measure of a fair price. Under Black-Scholes, this is the same for puts and calls when measured in volatility units. Thus, these drop out when using the call-put spread. Looking at either individually is more challenging because we need to make some kind of assumptions to either estimate the fair value or the value of the premiums. Throughout this paper, I will make the simplifying assumption that the option premium is constant through time. Thus, changes in the implied volatility should be predictive after controlling for other known predictors of returns.

44 35 To correct for the vol-of-vol premium, I sort on changes in the impliedrealized volatility spread instead of the level. This is essentially allowing stochastic volatility, but does not allow changes in the premium. While this may seem restrictive, changes in the premium predict that returns will go in the opposite direction as my prediction for put options. That is, if changes in the implied-realized spread are related to changes in the vol-of-vol premium, then we should expect this to be positively related to future returns, even for put options. However, these changes are negatively related to future returns when measured using put option prices. In summary, I sort stocks into portfolios based on the change in the impliedrealized volatility spread. These quintiles are formed into portfolios that are rebalanced monthly. However, I first must control for another factor. Failing to do so leads to spurious results, and is one of the main reasons that there is debate as to whether the call-put volatility spread premium is in fact related to informed trading. Under the informed trading hypothesis, it should be the case that we can see two distinct patterns. A run-up in call prices, relative to observed volatility, should signal higher future equity returns. However the same experiment for puts should yield the opposite result. However, previous work has been unable to show this dichotomy. An et. al. (2014) shows that sorting on changes in call-implied volatilities, we see the expected monotonic relationship: Firms with larger run-ups in call prices experience higher subsequent equity returns. However, if we sort portfolios on the changes in put-implied volatilities, we see that there is no statistically significant difference between the portfolios, even after controlling for observed volatility changes. Worse,

45 36 the trend is monotonic in the same direction as the call sorts. This casts doubt on whether the expected call pattern is due to informed trading at all, and indeed, I suggest that it is not. Figure 3: Implied volatility cycle around earnings announcements This figure plots the implied volatility of Apple stock options from August 2013 to August The implied volatility plotted is the mean of call and put implied volatility indexes calculated by IVolatility.com. Red vertical lines indicate the dates of quarterly earnings announcements for Apple Inc. When sorting on changes in implied volatility, especially without controlling for changes in realized volatility, one is essentially sorting on whether or not a firm has an earnings announcement within the expiration. That implied volatilities are cyclical around earnings announcements is easily observable from a cursory glance at a firm s time series. The last year of implied volatility for Apple stock is shown in the Figure 3 above, along with red lines indicating earnings announcement dates), while a more general trend has been shown rigorously in Dubinsky and Johannes (2006). The fact that we see positive relation between call price runups and future equity returns, then, is simply a restatement of Frazzini and Lamont (2007), which showed that the equity